Looks like the "learn it all end-to-end" deep learning hype is starting to fade as some people realize putting some actual prior knowledge in computer vision systems is helpfull.
http://alexgkendall.com/computer_vision/have_we_forgotten_about_geometry_in_computer_vision/
@vbarrielle
As Douglas Adams pointed out the absurdity of expecting deep learning to be the Ultimate Solution 3001:
"Deep Thought was so powerful that it had already deduced the existence of rice pudding and income tax before it was even turned on".